Why am I not mentioned in ChatGPT?

People type this when they have a site, a LinkedIn, maybe a book, and ChatGPT still talks around them. The Dutch query is the same job: why is my name not in the answer.

There is no published ChatGPT citation ranking. Nobody can sell you a position in it. What exists is a short list of operating failures, and they are cheap to test in order.

How to get cited by ChatGPT is the operating note. This page is the diagnosis. Work the list in order. Do not start with a GEO retainer.

Find your symptom first

The three failures look identical in a chat window and have nothing in common underneath.

What you see	Most likely cause	First thing to check
No trace of you at all	The Search door is closed, or the corpus never fetched the URL	robots.txt, then your access logs for OAI-SearchBot
Your URL appears as a source chip, no name in the prose	The page has no entity a machine can attach	Is a person or organisation named in the HTML, near the claim
Your framing or your number appears, uncredited	Your claim lives somewhere the engine credits instead, or it is unattributable	Where the canonical version of that sentence actually sits

Everything below is one of those three, in the order that costs you the least to fix.

You tested the wrong chat

ChatGPT’s training memory and ChatGPT Search are different doors.

Ask the question in Search, with sources visible, not in a private thread that is answering from training and from what it already knows about you. Use a fresh or temporary chat. A thread that has been talking to you for six months is not a neutral instrument; you may be reading your own history back.

Then log three outcomes as three different results, because their fixes are different:

• Named. The prose says your name or your organisation. This is the thing you actually wanted.
• Linked, not named. A source chip points at your URL and the sentence credits nobody. Retrieval worked; attribution did not.
• Used, not cited. Your framing, phrase, or number appears with no link and no name. Treat this as a lead, not proof: the wording may be common or independently produced.

“It knew the topic” is not any of the three. If you only ever test a memory chat, you will diagnose the wrong system for a year.

One run is not a result

The same prompt does not reliably return the same sources twice. Sources can vary by run, account, date, and phrasing.

So do not screenshot one answer and call it a finding. Write down five to ten prompts a real buyer would type, run each in a fresh Search chat, and record the outcome per prompt per run. You are looking for a pattern, not an anecdote.

Two things people skip:

• Language. Ask in Dutch and in English, and log them separately. Language can change the candidate set. If your market is NL, an English-only test tells you very little.
• The branded prompt. “Who is the leading X” is a vanity query nobody types. Test the question a customer asks the week before they buy.

The Search door is closed

OpenAI’s publisher FAQ says summaries and snippets in ChatGPT Search require you not to block OAI-SearchBot. That is eligibility. GPTBot is a separate training switch. ChatGPT-User is a person-triggered fetch. Mixing the three is how a robots.txt “fix” does nothing.

Fetch robots.txt on the live host. If OAI-SearchBot is disallowed on that URL, you are not in the Search corpus. Allow it on the pages you want quoted. OpenAI has said a Search-side robots change can take about 24 hours. Recheck then. Do not declare failure at minute ten.

Then check the doors that are not in robots.txt. A 403 from a bot rule, a login wall, a soft-404, a noindex you forgot, or a canonical pointing at a blocked URL are all the same closed door with a friendlier error page. Your access logs settle the first question: if no fetch from the Search crawler ever arrives, solve access and discovery before you rewrite the page.

The first HTML screen cannot be stolen

Give Search a clean capsule in the initial HTML. A claim buried in a client-rendered widget, paragraph eight, or a PDF is harder to extract and attribute. Often the first usable text is a nav or a manifesto.

Fetch the URL with no JavaScript. Highlight the first 80 words.

1. Is there a claim a stranger could repeat?
2. Is a person or organisation named (not “we”)?
3. Is the evidence on this URL?

If step one fails, you do not have a ChatGPT problem. You have an AEO opening problem. GEO is the system around the crawl. AEO is this URL.

Step two is the one that produces linked, not named. “We help founders” gives a machine nothing to attach. A full name, an organisation, and a date provide explicit entity and time cues.

Someone else owns your best sentence

This is the quiet one, and it can explain used, not cited results.

Your sharpest claim may sit on a platform: a LinkedIn post, a Medium piece, a Substack, a marketplace listing, a podcast show-notes page, a client’s case study with your work and their byline. The engine may credit that host. The sentence lives there.

The fix is unglamorous. Publish the canonical version on your own URL first, dated, with the evidence attached. Let the platform carry the short version and point home. If a claim only exists as a repost, you are asking to be summarised into somebody else’s brand.

Same failure inside your own site: if the claim lives on five near-identical pages, none of them is the obvious one to quote.

You are eligible and still unnamed

Eligibility is not a citation. Once the crawler can fetch you, you still have to win among other eligible pages.

Usual misses:

• The sentence is true everywhere. “AI is changing work” gives the engine little reason to carry your name.
• No date, no locator, no source on the same URL. The engine cannot verify you.
• The entity is a logo, not text. Machines attach names they can read.
• The page answers a topic instead of a question. Category pages and aggregators are better at topics than you are.

Write one unique, dated, sourced sentence that only you would put on that URL. Put your name next to it. Then run the prompt set again and log the three outcomes.

What this is not

It is not a secret prompt. Hidden instructions in a footer are not a strategy.

It is not FAQ schema. Google ended FAQ rich results on 2026-05-07. A fake FAQPage is a liability there and does not make ChatGPT name you.

It is not fame. Fame is not a technical requirement for eligibility or attribution, and it does not open a blocked door.

It is not a volume problem. Ten more commodity posts do not fix an entity that no machine can read.

The fail list, in order

1. Fresh Search chat, not memory. Write the prompt down.
2. Five to ten real buyer prompts, more than one run, both languages if you sell in both.
3. robots.txt: OAI-SearchBot allowed on this URL?
4. Access logs: has the Search crawler actually fetched it?
5. curl the HTML. Claim, name, and evidence in the first screen?
6. Where does the canonical version of that claim live: your URL, or a platform?
7. Wait a day after a robots change. Test again.
8. Log named, linked-not-named, used-not-cited.

Stop at the first failure and fix it before you spend on the next line. If 3 or 4 fails, do not move to line 5. If 5 or 6 fails, stop buying GEO theatre and fix the capsule. Then read how to get cited for the rest of the operating loop, and how answer engines select sources for what happens after the door is open.